[Zhipu's Coding Comeback: DeepSeek Forced a Shift, Users Chose the Path for Them] Beating AI News Flash: 36Kr reviewed the behind-the-scenes process of Zhipu's comeback through Coding. After DeepSeek R1 went viral during the 2025 Spring Festival, Zhipu's previously most relied-upon enterprise customization business took a direct hit. An employee estimated that nearly 30% of clients switched to DeepSeek at the time. To retain customers, Zhipu even offered very low discounts. In May, Zhipu held a critical strategy meeting. The company abandoned its original approach of developing vertical models for text, multimodal, and programming separately, deciding instead to integrate Reasoning, Coding, and Agentic capabilities into a single large model. This sudden pivot delayed the originally planned April release of GLM-4.5 to July. Interestingly, Zhipu did not initially prioritize Coding. Among the three capabilities, there was no clear hierarchy at first. What truly pushed the model toward Coding was user feedback. During development, "improving R&D efficiency" was consistently the most common demand from customers. After GLM-4.5 launched, more and more enterprises requested to integrate it into programmer workflows, and developers began using it as a substitute for Claude. Coding quickly became the most well-received capability of GLM-4.5. After this, Zhipu began narrowing its focus. The Coding and text teams currently have around 100 members each, while the multimodal training team has been reduced to just 10 people. Training has increasingly emphasized data, fine-tuning, and real customer feedback. When uncertain about technical directions, they let several research teams run A/B tests and chose the one with the best results. Before releasing a model, they tested general capabilities while also having the delivery team conduct another round of testing in real-world scenarios. Following GLM-4.5, Zhipu launched the Coding Plan and enterprise edition, making Coding one of the most important capabilities in the GLM series. By the time GLM-5.2 was released, both the model's reputation and API revenue soared. Zhipu's ARR jumped from $500–600 million in May to $1 billion by July. Zhipu didn't initially bet on Coding. DeepSeek first disrupted its comfort zone, and customers then chose a new direction for it. [Original Article Link]